The same DNA, a different context

Google DeepMind introduced AlphaGenome 2 on November 11. The model predicts how a stretch of DNA, or a change to it, might affect molecular activity in a specified cell type or state. A change that has little predicted effect in one cell type could matter in another, because different cells use different parts of the same genetic instructions.

The earlier AlphaGenome produced a fixed set of predictions across many tissues and cell lines. In the new version, a researcher supplies the context along with the DNA. They can choose from the reference catalogue or, as DeepMind describes it, use a profile derived from their own single-cell experiment.

Check which experiments were actually measured

A researcher can ask for an estimate of gene activity or how accessible a stretch of DNA is. The software can also estimate a result for an experiment that was never performed for that cell type. Such a result can help decide what to measure next. It is not a measurement.

The documentation labels whether the corresponding experiment was measured in the training data, inferred from other informative measurements, or treated as exploratory. Even in the first category, the value returned by the model is still a prediction. DeepMind warns that some exploratory outputs may be misleading because the available data do not constrain them well.

A tool for research

DeepMind says AlphaGenome 2 is available through its API for non-commercial research, with a separate private preview for commercial research. The release does not establish that it can diagnose a patient or that a predicted DNA effect causes a disease. A useful next step is to check how well the model’s predictions hold up in new experiments, particularly where little was measured in training.

Sources

  1. Google DeepMind: Introducing AlphaGenome 2Dated release, change from fixed outputs to condition-specific predictions, and stated access.
  2. AlphaGenome documentation: Choosing a conditionHow to choose a tissue, cell type or user-supplied cell condition for a prediction.
  3. AlphaGenome documentation: Navigating imputed tracksDefinitions of measured-in-training, supported and exploratory output categories; limitations of inference.